[kklimuk]

🤖 docx-cli triage

What was wrong

The reader skipped every run-level <w:sdt> (inline content control), so the text inside one never appeared. That covers a Word form's checkbox glyph and any Plain Text or Dropdown field value. For the issue's paragraph, docx read printed Approved <!-- p1 -->, and docx find de.docx "☐" printed no matches (exit 0). wc didn't count the text either.

What changed

• An inline <w:sdt> is now a run-bearing wrapper: it is in RUN_BEARING_WRAPPER_TAGS, and wrapperContentNode(sdt) returns its <w:sdtContent>. The reader (walkInlineContentControl) and every XML offset walker step into the same node, so find, replace, comments add and span edit offsets agree.
• A checkbox Word clicked under tracking (the ins/del glyph pair) is still one checkboxToggle tcN, and accept/reject keep the glyph and w14:checked in step. Other revisions inside a control now get ordinary tcNs, and accepting or rejecting one changes only the content inside that control.
• When a span cuts across a control's edge, each piece stays inside its own copy of the control. Each copy keeps the control's properties except <w:id>, so no two controls share an id.
• A --text rewrite that repeats the control's text keeps that text inside the control, the same way it already keeps hyperlinks (keptWrapper). A checkbox stays a clickable checkbox.
• Docs: REFERENCE.md (content controls bullet), plus the nested CLAUDE.md files for core, ast, task-list and track-changes.
• Fixture: transparent-wrappers.docx gains a third paragraph with a checkbox and a Plain Text field.
• Tests: 7 new Checkbox content controls (w14:checkbox) are invisible in read and cannot be set #58 tests in tests/cli/invariants.test.ts. Two existing tests that asserted the old skip are rewritten for the new behaviour.

Verification

• Before: read → Approved <!-- docx:p p1 raw -->; find "☐" → no matches.
• After: read → ☐ Approved <!-- docx:p p1 raw -->; find "☐" → p1:0-1; wc counts the glyph.
• Also checked by hand: replace inside a field, replace across a field's edge, comments, footnotes, --bold, --text with and without --track. validate is clean after each.
• bun run check ✅ bun run test:unit ✅ (2107 tests in the worktree; 2137 in the pre-commit hook run)
• /simplify (single pass, no Agent tool): merged the two copies of the checkbox-toggle registration into one registerCheckboxToggle helper. /code-review xhigh: the forked run reviewed the wrong scope (commit feat(images)!: refuse HEIC/HEIF input and drop the LGPL libheif dependency #56 in the main checkout), so I didn't apply its output. I did a manual adversarial pass on this diff instead and found no new bugs (details below).

Decisions for you

• A verb to toggle a form checkbox. docx tasks check/uncheck --at pN handles task-list items only. A form checkbox (Yes/No in a table cell, or several boxes in one paragraph) needs an address for each box, either a new id (e.g. cbN) or a span such as pN:S. Until then, replace ☐ ☒ changes the glyph but leaves w14:checked alone. REFERENCE.md says so.
• Should read add a hint for checkboxes? For example <!-- docx:checkbox pN:0 checked=no -->, which would tell an agent the glyph is a control and not text. I left it out until the verb question above is settled.
• Bugs that existed before this PR and now also affect content controls (they already affect fldSimple/smartTag/hyperlinks):
  • A tracked whole-paragraph replace (--track, or --markdown, over a paragraph with pending revisions) leaves wrapper text undeleted. This doubles a glyph, e.g. ☒☒ Buy bread.
  • A tracked span that cuts across a wrapper's edge drops the cut inner text with no <w:del>, so it can't be restored by reject. Before this PR, fldSimple already did this (Today is -05-05).
  • A --markdown whole-paragraph edit flattens the control to plain text.
• Changed behaviour in one case. An untracked whole-paragraph --text edit over a checkbox toggle that is still pending now resolves the toggle and prints the Whole-paragraph edit --text --track silently resolves other authors' pending changes #17 warning. That is how other pending text behaves. Before, the control was carried through unseen.

Fixes #58

Action needed: approve/merge — review the "Decisions for you" checklist (above all, whether to add a checkbox-toggle verb in this PR or a follow-up), then merge.

---

Generated by Claude Code